Tether’s data and AI unit QVAC announced a cross-platform BitNet LoRA fine-tuning framework on March 17, integrating the system into QVAC Fabric. The company said it is the first framework of its kind to support Microsoft’s BitNet 1-bit LLM architecture, with a focus on moving model fine-tuning from data-center hardware to consumer devices such as smartphones and laptops.
The release targets billion-parameter models running locally. According to QVAC, the combination of BitNet’s 1-bit design and its own optimization cuts memory use and compute demand enough to make on-device training practical on mainstream hardware. Phones are part of that plan. So are notebooks.
Support spans desktop chips and mobile GPUs
In its announcement, QVAC said the framework works with NVIDIA GPUs and also supports Intel, AMD, and Apple M-series chips, along with mobile graphics on Adreno, Mali, and Apple Bionic. Tether said edge devices using the framework can handle models 2x larger than traditional Q4-quantized models, pointing to memory efficiency as the main advantage.
That brings LoRA fine-tuning to a wider range of endpoints. Personalized training that usually sits on servers is being compressed into the limits of local hardware, which is a central part of QVAC Fabric’s pitch.
Benchmark numbers from Samsung S25 and iPhone 16
Tether’s engineering team also shared test results from recent smartphones. For a 125 million-parameter model, fine-tuning on a dataset of 300 biomedical documents took about 10 minutes on a Samsung S25.
For a 1 billion-parameter model, the same task took 1 hour and 18 minutes on the Samsung S25 and 1 hour and 45 minutes on the iPhone 16. The team also said it managed to run fine-tuning for a 13 billion-parameter model on an iPhone 16, describing it as a push toward the physical limits of mobile hardware.
Local-first design avoids sending data to the cloud
Tether CEO Paolo Ardoino has argued that if an AI product requires an API key, it does not truly belong to the user. QVAC’s position is built around a local-first approach.
Under that model, users can fine-tune AI systems on emails, notes, and messages stored on their own devices without uploading the data to cloud servers. The point is direct control over data handling. For companies, that reduces concerns around sensitive information leaving local environments; for developers, it lowers dependence on centralized compute and platform access.
QVAC Fabric LLM has been released as open-source software under the Apache 2.0 license, and preselected adapters are available on Hugging Face. Tether is pushing an edge-compute route in which training and personalization happen on end-user devices rather than only in the cloud.